Perturbation bounds on the polar decomposition (Q913900): Difference between revisions

Let A be a real \(n\times n\)-matrix with a polar decomposition \(A=MH\), \(H=H^ T\) positive semi-definite, M orthogonal. Compare polar decompositions \((M+\Delta M)(H+\Delta H)=A+\Delta A\) for perturbations \(A+\Delta A\) with \(MH=A\). If \(\sigma_ 1\geq...\geq \sigma_ n\geq 0\) denote the singular values of A, and \(\sigma_{n-1}>0\) then \(\| \Delta H\|_ F\leq \sqrt{2}\| \Delta A\|_ F\) if \(\sqrt{2}\| \Delta A\|_ F<\sigma_ n+\sigma_{n-1}\). If \(\sigma_ n>0\) and \(\| A\|_ 2<\sigma_ n\) then \[ \| \Delta M\|_ F\leq - \sqrt{2} \ell n(1-\frac{\sqrt{2}\| \Delta A\|_ F}{\sigma_ n+\sigma_{n-1}}); \] here \(\| \|_ F\) denotes the Frobenius norm, and \(\| \|_ 2\) is the matrix norm induced by the Euclidean norm. The estimate for \(\Delta\) H is also valid in the complex case, but the estimate for \(\Delta\) M, which is much better than bounds previously known in literature if \(\sigma_{n-1}\gg \sigma_ n\), is only valid in the real case. For the estimate for \(\Delta\) M applications to the orthogonal Procrustes problem and to the kinematics of a rigid body are sketched.